Innovation Monitor: Could AlphaFold spark a bioscience revolution?

Innovation Monitor: Could AlphaFold spark a bioscience revolution?

View this email in your browser

Welcome to this week’s Innovation Monitor.

We’re delving into deep tech and science, and more specifically, the world of proteins. When we look back at the past several decades in AI, two moments stand out: the 2012 ImageNet moment, when an AI model from the University of Toronto outperformed other teams in image classification by a significant margin, kicking off the AI boom of the 2010s, and late last November, when DeepMind got international attention for its unprecedented performance at the Critical Assessment of protein Structure Prediction (CASP) competition.

At the time, DeepMind’s protein-folding-prediction model, AlphaFold 2, received a score that was “informally considered to be competitive with results obtained from experimental methods.” These methods typically take years to complete per protein structure, while AlphaFold spends just days for its predictions.

We’ve seen first-hand the cascade of advancements around CV since 2012 — from the way your phone auto-tags certain photos, to self-driving cars, to agricultural bots, to Orwellian camera networks around the world. It’s brought on as many ethical and governance issues as societal benefits.

AlphaFold’s accomplishment is harder to judge this early — like so much of deep tech — but it’s not a leap to say it surpasses the ImageNet milestone. This week we’re going to look at why that is, starting with what protein folding is, and why it’s so hard to get right. We’ve all seen the incredible progress of mRNA vaccines (which this newsletter recently covered in detail) and the open-source nature of AlphaFold could unlock new, collaborative discoveries that can improve our lives.

Thank you for reading, and as always, if you were forwarded this email, you can easily sign up here!

All best,
Erica Matsumoto What is protein folding? Why is it so difficult? With the help of AlphaFold, DeepMind recently released “the most comprehensive map of human proteins,” consisting of the predictions of some 350k protein structures, including ~250k that were previously unknown, according to a NY Times feature.

Scientists are likening the potential impact of the data to that of the Human Genome Project (even two decades after that, the project’s massive impact isn’t so apparent unless you’re reading up): “Most significantly, the [DeepMind] release includes predictions for 98% of all human proteins, around 20,000 different structures…. it isn’t the first public dataset of human proteins, but it is the most comprehensive and accurate.”

So what’s protein folding and why is it so difficult? As Nature wrote last November: “Proteins are the building blocks of life, responsible for most of what happens inside cells. How a protein works and what it does is determined by its 3D shape — ‘structure is function’ is an axiom of molecular biology…. Accurately predict protein structures would vastly accelerate efforts to understand the building blocks of cells and enable quicker and more advanced drug discovery.”

Since the 1950s, X-ray crystallography has been helpful in discerning protein structures. In the past decade, cryogenic electron microscopy has proven to be a powerful technique as well. But computer techniques have performed poorly until the latter half of this decade (“lofty claims for methods in published papers tended to disintegrate when other scientists applied them to other proteins,” according to the Nature piece).

But like mentioned above, these methods take years — and “tens or hundreds of thousands of dollars per protein structure,” according to a DeepMind blog. Hence why AlphaFold’s accomplishment was praised by biologists. “This will change medicine. It will change research. It will change bioengineering. It will change everything,” Andrei Lupas, an evolutionary biologist at the Max Planck Institute for Developmental Biology, told Nature back in November.

What is the impact of AlphaFold and its database? DeepMind co-founder and CEO Demis Hassabis has been interested in the protein folding problem for over 20 years, according to an interview with MIT Tech Review: “It’s been a huge project for us. I would say this is the biggest thing we’ve done so far. And it’s the most exciting in a way, because it should have the biggest impact in the world outside of AI.”

For a great brief on AlphaFold’s potential for structural biology, we recommend going through CIMR research associate Tristan Croll’s tweet storm. The gist is, AlphaFold is a “spookily accurate” supplement to a biologist’s work.

And similar to how everybody rushed to get an OpenAI beta key once GPT-3 landed, scientists all over the world are having fun exploring DeepMind’s database.

That’s where the power lies — the open source nature of the database. And it’s just the beginning. DeepMind partnered with the European Bioinformatics Institute for the database, which will eventually get millions of AlphaFold structure predictions out into the open. The institute released an excellent breakdown of how the database will affect the scientific community.

“In drug discovery, the use of 3D models can help understand why a certain drug compound is an inhibitor but not a related compound, or why certain proteins are ‘druggable’ while others are not. The models will accelerate research efforts to identify new candidate drugs and even drug targets as a predicted model will now often be available to start looking for druggable sites, to kick-start structure determination, to design targeted ligand or fragment screens, etc. Virtual screening techniques may also suggest new uses for old drugs against targets for which no structure was previously available.” Wrapping up Technology and biology are intertwined more than ever. As a consequence, biology has become vastly more distributed and accessible (actually, you can go ahead and give a quick look through the structure prediction database right now). The convergence of AI, biology, and high-throughput screening is creating something more than the sum of its parts, and eventually it will transform processes around “health and disease, with applications in biotechnology, medicine, agriculture, food science and bioengineering.” This Week in Innovation History

August 3rd, 1993: Apple Introduces the Newton

On this day, Apple launched one of the world’s first Personal Digital Assistants (PDAs) which are a much more accurate description of the devices we now carry in our pockets versus a “phone”. While this was a precursor to our smartphones, one of the most important innovations to come from the device was the implementation of the ARM Microprocessor. Apple had invested heavily into a British company to help bring these powerful, low-energy consumption chips to push the limits of a portable device. This chip architecture continues to power Apple portable devices and has “saved Apple…twice”.

This email was sent to <<Email Address>>
why did I get this? unsubscribe from this list update subscription preferences
NYC Media Lab · 370 Jay Street, 3rd floor · Brooklyn, New York 11201 · USA



NYC Media Lab connects university researchers and NYC’s media tech companies to create a new community of digital media & tech innovators in New York City.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
NYC Media Lab

NYC Media Lab connects university researchers and NYC’s media tech companies to create a new community of digital media & tech innovators in New York City.